import torch
@torch.compile(backend="eager", fullgraph=True)
def f(x):
   y = x.clone()
   def g():
      return y + 1
   return g()

g = f(torch.randn(4))

running the above with TORCH_LOGS=+graph_code gives the following graph:

def forward(self, L_x_: "f32[4]"):
    l_x_ = L_x_
    y: "f32[4]" = l_x_.clone();  l_x_ = None
    add: "f32[4]" = y + 1
    return (y, add)

Dynamo doesn't need to return y because it is not accessible from the outside of the compiled function. The reason why it does is:

• every time Dynamo captures a Tensor, it writes a new side effect
• If any Tensor participates in a live side effect, then Dynamo must return it as an output of the graph
• This side effect never gets pruned, so Dynamo returns it as an output of the graph

This is potentially a performance issue for models that do a lot of capturing of variables

cc @ezyang @msaroufim @bdhirsh @anijain2305 @chauhang @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @kadeng